YOLO Object Detection Plugin for DeepStream 2.0

Hi all, we have released a new sample plugin for DeepStream 2.0 performing YOLO (You Only Look Once) object detection, accelerated with TensorRT. Supported models include YOLO v2 & v3.

Find the sources here: https://github.com/vat-nvidia/deepstream-plugins

Hello dusty_nv,
I am very interesterd in your plugin. Actually my Nvidia Jetson TX2 is running Deepstream sdk 1.5. I see your puglin require a Deepstream sdk 2.0. On nvidia deepstream dowload page only 1.5 version is avaible for Jetson TX2, is there a Deepstream 2.0 sdk version for Jetson TX2 ?
Otherwise does your plugin will work on Jetson TX2 DeepStream 1.5 sdk ?

Thnxxx in advance
Marco Gonnelli

I run the YOLO v3 plugin in the deepstream app using Tesla P100. The video file source is sample_720.mp4 and the number of sources is one. The average fps is only 18. When I set the number of sources 10, the average fps becomes 4. Why it is so slow?

@m.gonnelli Currently this plugin is only tested for DS 2.0 but you should be able to adapt the core of the plugin for DS 1.5 as well.

@lihongbo14 Currently V3 performance is CPU bound due to a few post processing functions. There is more room for improvement in this regard. This sample was meant to serve as an example to help users get started with using yolo v3 in DeepStream/TRT.

Building the TRT-yolo app for the plugin.
Used TensorRT4.0 with Opencv 3.4

Got the following error for the makefile in sources/apps/TRT-yolo

…/…/gst-yoloplugin/yoloplugin_lib/build/yoloplugin_lib.o: In function YoloPluginProcess': yoloplugin_lib.cpp:(.text+0xf00): undefined reference to cv::dnn::experimental_dnn_v3::blobFromImages(std::vector<cv::Mat, std::allocatorcv::Mat > const&, double, cv::Size_, cv::Scalar_ const&, bool, bool)’
collect2: error: ld returned 1 exit status
Makefile:49: recipe for target ‘TRT-yolo-app’ failed

Error is with the yoloplugin_lib referencing to opencv dnn module. But the Makefile contains the -lopencv_dnn flag to build with.

Any idea to implement yolov3 or yolov2(not tiny-yolo) at real-time fps(30 fps) in jetson-tx2 with tensor-rt?

Were u able to resolve this issue as am facing the same one.